13 mecanica de fluidos 12
Post on 05-Apr-2018
222 Views
Preview:
TRANSCRIPT
-
8/2/2019 13 Mecanica de Fluidos 12
1/20
TESTS OF PERFORMANCE OF THE BOUNDARY ELEMENT DUAL
RECIPROCITY METHODMULTI-DOMAIN APPROACH FOR 3D
PROBLEMS
Bruno Natalinia
and Viktor Popovb,
aFacultad deIngeniera, Universidad Nacional del Nordeste, Av. Las Heras 727, 3500 Resistencia,
Argentina, bnatalini_2000@yahoo.com.ar, http://ing.unne.edu.ar/tunel/pagina%20tunel.htmbWessex Institute of Technology, Environmental Fluid Mechanics, Ashurst Lodge, Southampton, SO40
7AA, United Kingdom, viktor@wessex.ac.uk, http://www.wessex.ac.uk
Keywords: Boundary element method; DRM-MD; Optimal implementation; Radial basis
functions; 3D problems.
Abstract. In this paper, aspects regarding implementation of the boundary element dual reciprocity
methodmulti-domain approach (DRM-MD), in respect to 3D problems are reviewed. Results of
numerical tests on a 3D advectiondiffusion problem with non-uniform velocity field are presented.
The sensitivity of the accuracy and stability of the codes to the continuity of the elements, scaling,
internal DRM nodes and mesh refinement have been tested. The results show that scaling is essential
and that mesh refinement and/or internal DRM nodes improve the accuracy when the non-
homogeneous term of the governing equation becomes dominant. The computer code implemented
with discontinuous elements offers higher accuracy, especially for advection dominant transport, but is
much slower than the computer code with continuous elements. At the present stage, the discontinuous
element code has the advantage of flexibility since it can solve non-homogeneous domains.
-
8/2/2019 13 Mecanica de Fluidos 12
2/20
1 INTRODUCTION
The dual reciprocity method (DRM), which was introduced by Nardini and Brebbia
(1983), is acknowledged to be one of the most effective boundary element method (BEM)
techniques for transforming domain integrals into boundary integrals.The accuracy and stability of the DRM is strongly dependent on the function used in the
DRM approximation, for which a variety of interpolation functions can be used. Until the
1990s simple ad hoc expansions such as 1+R, whereR is the distance between a collocation
point and a field point, were preferred by users. With the application of the theory of radial
basis functions (RBFs) in the context of the DRM the procedure gained a solid mathematical
foundation. The origin of RBFs can be traced back when they were applied to geophysical
data interpolation. In the 1990s, Golberg and Chen (1994, 1996) introduced the theory of
RBFs in the DRM field and demonstrated that the DRM converges if RBFs are used as
interpolation function. They pointed out that the early success of 1+R as approximation
function in the DRM is due to the fact that it belongs to RBFs.A characteristic feature of the DRM is that it uses a set of internal points to improve its
accuracy. The accuracy of the method is sensitive to both the number and distribution of the
DRM points. These two factors constitute its main drawbacks, as there are no standard criteria
to deal with them. There have been some attempts in using adaptive techniques (Schclar,
1993; Rodrguez and Power, 2002) in order to approach this problem.
The DRM has been demonstrated to be a general and reliable procedure. However, as
many of the RBFs used are globally supported the matrix of the resulting system of equations
is dense and frequently ill conditioned, when applied to large problems. This makes the
method computationally expensive and sometimes unstable. There are two ways to avoid
these difficulties: by using compactly supported RBFs (CS-RBFs) or by using domain
decomposition.Positive-definite CS-RBFs, which are locally supported, have been explicitly constructed
and applied to multivariate safe reconstruction in mid-1990s by Schaback (1995), Wendland
(1995) and Wu (1995). Since then, much effort has been focused on building efficient
algorithms using CS-RBFs, as a result of which new functions have been proposed
(Wendland, 1995; Buhmann, 2001). The implementation of these functions leads to a sparse
system of equation that is free of the problems mentioned above, due to the local support.
However, the current stateof-the-art CS-RBFs faces two main difficulties: (a) their accuracy
and efficiency depend on the scale of the support, with the scale being uncertain, and (b) the
convergence rate of CS-RBFs is low.
Domain decomposition is a technique that is commonly used in the BEM when the domain
is piecewise homogeneous. After applying the numerical formulation in every subdomain, the
final system of equations is obtained by means of a set of matching conditions in the
interfaces between subdomains. The resulting system of equations is not dense, and the
sparsity of the system increases with the number of subdomains. Popov and Power
implemented a scheme using domain subdivision in conjunction with the DRM to avoid
domain integration and called it the dual reciprocity methodmulti-domain approach (DRM-
MD). The initial problem solved using this formulation was the flow of a mixture of gases
through a porous media (Popov, Power and Baldasano, 1998; Popov and Power, 1999a,
2000). The DRM-MD has also been applied to linear and nonlinear advectiondiffusion
problems (Popov and Power, 1999b), driven cavity flow of NavierStokes equations (Florez
and Power, 2002a) and of non-Newtonian fluids (Florez and Power, 2002b), and the flow ofpolymers inside mixers with complex geometries (Florez, 2001). DRM-MD does not suffer
the two main problems related to standard DRM; the systems of equations produced by DRM-
-
8/2/2019 13 Mecanica de Fluidos 12
3/20
MD are sparse and well conditioned, and the number and position of DRM nodes is usually
not critical, since small sub-domains usually require no or few interior DRM nodes.
Most of the reported results of the DRM-MD in the literature refer to 2D applications. As
the method was extended to 3D problems in 2004 (Natalini and Popov, 2004), presently a fewfacts are known about 3D implementation, though this is an issue that is far from being
settled. For instance, Natalini and Popov (2004, 2006) tested 10 different DRM
approximation functions, five globally and five compactly supported RBFs, in codes that
solve Poisson and advectiondiffusion problems. The highest accuracies were obtained by
using compactly supported RBFs. However, a suitable size of the support must be known a
priory and there is no rigorous guideline to choose it. It is known that a small size of the
support guarantees a safe recovery, but a large one reduces the error at expense of stability.
Therefore a balance must be reached between these two factors. This topic has been discussed
within the framework of multiscattered interpolation theory by Floater and Iske (1996) and
Schaback (1995) among others. Floater and Iske proposed a strategy but it is to be used in a
hierarchical scheme for smoothly interpolating scattered data, which cannot be applied in thenumerical models presented here or the ones of Natalini and Popov (2004, 2006).
Consequently Natalini and Popov concluded that the augmented thin plate splines (ATPS)
appear to be the best choice at the moment, as ATPS produced one of the most accurate and
certainly the most consistent results without introducing any additional parameter.
On the other hand, Peratta and Popov (2006) applied the DRM-MD in a hybrid formulation
to large 3D problems of transport in fractured porous media. Based on comprehensive
research by Portapila and Power (2001, 2005) on the application of iterative solvers in 2D
DRM-MD codes, they used an iterative solver combined with a MC64 preconditioner. The
preconditioner showed to be essential since it reduces the CPU time required by the solver by
several orders of magnitude.In this paper, the sensitivity of the method to factors such as continuity of the elements,
scaling, mesh refinement and number of internal nodes is tested for an advection-diffusion
problem with variable velocity field, in order to have a better insight into the 3D
implementation of the DRM-MD.
2 THE DUAL RECIPROCITY METHOD
Let us consider the following Poisson equation, with the transport coefficient equal to one
)(2 xu = b(x) (1)
where )(xu is a scalar field (potential function), )(xb the non-homogeneous term, and x the
position vector in the domain with components ei.
Given a point x belonging to a domain , which is enclosed by a contour, the Greenintegral representation formula for(1) gives the value ofu at x in terms of integral equations
involving the fundamental solution of the Laplace equation:
=+ yyy dbudquduqu )(),()(),()(),()()(***
yyxyyxyyxxx (2)
Here, u*(x,y) is the fundamental solution of the Laplace equation given by
r
u1
4
1),(
*
=yx (3)
for 3D problems, where ris the distance from the point of application of the concentrated unit
-
8/2/2019 13 Mecanica de Fluidos 12
4/20
source to any other point under consideration, i.e. r = |x-y|, q(y) = u(y)/n and q*(x,y) =u*(x,y)/n and n is the unit normal to the boundary of the sub-domain. Notice that in Eq. (2)all the integrals are over the boundary of the domain except for the one corresponding to the
term b(y), which represents the sum of the non-homogeneous terms. The constant (x) hasvalues between 1 and 0, being equal to 1/2 for smooth boundaries.
To express the domain integral in (2) in terms of equivalent boundary integrals, the DRM
approximation is introduced. The basic idea is to expand the b(y) term using approximation
functions, i.e.:
+
=
=IJ
k
kkfbb
1
),(~
)( zyy (4)
The functions f(y,zk) are approximation functions, which depend only on the geometry of
the problem, and the constants k are unknown coefficients. The approximation is done at
(J+I) nodes, withJboundary nodes around the boundary of the domain andInodes inside thedomain.
Once the DRM approximation of the non-homogeneous term, b(y), is implemented, the
domain integral can be recast in terms of a series of surface integrals, and one finally arrives
at a boundary only integral representation formula
+ yy dquduqu )(),()(),()()(**
yyxyyxxx
+
=
+
IJ
ky
ky
kkk dquduqu
1
** ),(),(),(),(),()( zyyxzyyxzxx (5)
For the numerical solution of the problem, the contour is discretized in j elements and
the density of the integrals in the above equation is defined in terms of nodal values by means
of interpolation functions. After application of collocation technique (5) can be written in
terms of four matrices, H, G, U and Q which depend only on the geometry of the problem
qu )( QGUHGH = (6)
In (6) the vector is unknown but it can be expressed as b1
= F , yielding
bqu-1
)( FQGUHGH = (7)
3 THE DUAL RECIPROCITY METHODMULTI-DOMAIN APPROACH
The domain discretization in the BEM is usually used when there are few parts of the
domain with different properties. In that case the domain decomposition is often used, in
which the original domain is divided into subregions, and on each of them the full integral
representation formula are applied. A case of a domain, which is subdivided into four sub-
domains, is shown in Figure 1. Though in Figure 1 a 2D domain is considered for the sake of
simplicity, the conclusions can be extended to 3D cases as well.
Matching conditions for a potential problem establish that at every node at the interface:
(a) the value of the potential is the same for both subdomains. For instance, at node F ofFigure 1
-
8/2/2019 13 Mecanica de Fluidos 12
5/20
)()( 2112 FF
FF
uu xx = (8)
(b) the physical flux is the same for both subdomains
),(),( 212121121212FFFFFF ququ = (9)
where the form of the function F depends on the physical problem under consideration.
q
q12
q43
q23
14
1
2 3
4
A
B
C
D
EF
12qF
E
E E
E
Fq21
Figure 1: Example of subdivision of a domain into sub-domains.
While the BEM matrices that arise in the single domain formulation are fully populated,
the sub-region formulation leads to blocks banded matrix systems with one block for each
sub-region and overlaps between blocks when subregions have a common interface. Eq. (10)
represents the structure of the system of equations in matrix form that corresponds to domain
sub-division shown in Figure 1.
=
4
3
2
1
4
34
3
23
14
2
12
1
44341
34332
23221
14121
00000
00000
00000
00000
m
m
m
m
p
p
p
p
p
p
p
p
AAA
AAA
AAA
AAA
DE
CE
AE
BE
DEAE
DECE
CEBE
AEBE
(10)
Aj represents the influence coefficients obtained by integration over the external boundary
that bounds the sub-domainj andpi represents the unknown potentials uj or derivatives qj at
the nodes on this part of the boundary. For example, for the sub-domain 1 the external part of
the boundary is given with the curve from A to B.i
klA represents the influence coefficients
obtained by integration over the interface of the sub-domains kand landiklp represents the
unknown potentials and derivatives at the nodes on the interface. When considering nodes on
the interface several different situations may occur, of which only the most characteristic two
will be explained in this text. The first one will be analysed using the node F on the interface
between sub-domains 1 and 2. In this node there are four unknowns, two potentials and two
-
8/2/2019 13 Mecanica de Fluidos 12
6/20
normal derivatives. Two equations can be written collocating from the F node, one for the
sub-domain 1 and the other for the sub-domain 2. Using Eqs. (8) and (9) the contribution of
this node towards a closed system of equations is achieved. The situation with node E is more
complicated as this node is shared between four sub-domains. In each sub-domain there willbe three unknowns, two derivatives and one potential, which would overall make 12
unknowns. However, as the potential is unique in this node, using Eq. (8) three of the
unknowns are eliminated reducing the number of unknowns to nine. Further, by using Eq. (9)
the number of unknowns will be reduced to five, that is, four normal derivatives, for example
the ones shown in Figure 1, and the potential. With four equations that can be written
collocating from the node into each of the sub-domains, the contribution of this node towards
closed system of equations is not yet achieved and unless the medium is homogeneous and
the line/s BED or/and AEC are smooth at the node E, this node would need to be
converted to a discontinuous node in order that a closed system of equations is achieved.
Node E when discontinuous will have four freedom nodes instead, moved for a small distance
from the location of node E on the lines AEC, in the direction of A and C nodes, and on theline BED, in the direction of B and D nodes. In each of the new freedom nodes a situation
equivalent to the situation in node F will appear. When the medium is homogeneous and the
intersection lines are smooth in E, it can be shown that q12=q43 and q23=q14, reducing the
number of unknowns to three, making the final system of the equations over-determined.
Therefore, when continuous, node E may have three degrees of freedom, contributing towards
an over-determined system, or four degrees of freedom contributing towards a closed system
of algebraic equations, depending on whether both or just one of the BED and AEC lines
are smooth in the node E. Node E can be continuous if all of the sub-domains 14 are with the
same properties, or two by two of the neighbouring domains are of same properties, i.e., 12
and 34, or, 14 and 23. In any other combination node E must be discontinuous, whichproduces eight degrees of freedom, contributing towards a closed system of algebraic
equations. Similar analysis could be applied to nodes which are shared between three or more
than four sub-domains in 2D or 3D.
Next, let us define degree of overdetermination of acontinuous node, vO , as
vO = eqN - unN (11)
where eqN is the number of equations introduced by the node and unN the number of
unknowns at the node.
Provided all the subdomains around the node have the same properties, the value of eqN
and unN can be calculated as
eqN = subN + coN (12)
unN = 1 + inN (13)
where subN is the number of subdomains around the node, coN the number of
independent conditions of collinearity (in 2-D problems) or coplanarity (in 3-D problems) that
exist in respect to interfaces joining the node and inN the number of interfaces joining the
node.
4 MODEL FORMULATION
In this paper, results on problems governed by the advectiondiffusion equation
-
8/2/2019 13 Mecanica de Fluidos 12
7/20
02 = kuuVuDrr
(14)
are presented, whereD is the coefficient of diffusion, Vr
is the vector of flow velocity and kis
the reaction constant. Eq. (14) can be written as a non-homogeneous Laplace equationyielding a b term equal to ( )kuuV
D+
rr1. After applying the DRM formulation (7), the
resulting system of equations becomes
+
+
+
=
u
uV
uV
uVqu
321k
eeeD 3eee
21
1 1)( FQGUHGH (15)
where1e
V ,2e
V and3e
V are diagonal matrices containing the flow velocity components in the
directions e1, e2 and e3, respectively.
Replacing the partial derivatives by
uu 1
=
F
F
ii ee(16)
and denoting S the matrix1
)( FQGUH yields
+
+
+
=
uuVuVuVqu
321k
eeeD 3eee
11
2
1
1
FF
FF
FFS
GH (17)
Denoting as T the following matrix:
+
+
11
2
1
1
FF
FF
FFS
3
eeeeeeD 321
VVV and
reordering (17) produces
0=
qu G
S-T-H kD
(18)
5 COMPUTATIONAL IMPLEMENTATION
The implementation of the DRM-MD can be made in a variety of arrangements regarding
factors such as domain subdivision, continuity of elements, shape functions, etc. The results
presented in this paper were produced by two codes that basically differ in the continuity of
the boundary elements. Apart from this, both codes require that the domain be subdivided intetrahedral subdomains, being every side of every tetrahedron a triangular boundary element
which geometry is described by a quadratic shape function; consequently, the geometry of
every tetrahedron is described by ten geometrical nodes (Figure 2). This makes the codes very
flexible to model complicated geometries. Quadratic shape functions have been also used to
approximate the potentials and derivatives. The matrices arising from the model are stored in
sparse format. This feature together with the use of iterative solvers makes it possible to run a
wide range of problems on a PC.
The main complications arising when using multidomain BEM approaches are due to mesh
generation and assembly of equations. The problem of mesh generation is relatively easy to
solve since a variety of commercial mesh generators are available. The assembly is more
complicated in the subdomain BEM methods since potentials as well as derivatives need to beassembled as described in Section 3. For this purpose an assembling algorithm has been
-
8/2/2019 13 Mecanica de Fluidos 12
8/20
designed that is general and easily adaptable to problems with different governing equations.
The algorithm has been described byNatalini (2005).
Both codes use ATPS as DRM approximation function.
5.1 CODE A: a code using discontinuous boundary elements
CODE A is the same used by Natalini and Popov (2004, 2006) to test different RBFs,
which is to the authors knowledge the first ever reported 3D DRM-MD code. A modified
version of this code was produced by Peratta and Popov (2006) to simulate transport
phenomena in a waste repository. This code uses discontinuous triangular elements, following
the treatment proposed by Do Rgo Silva (1994), who described in detail the BEM
implementation. Over every triangular element, six nodes of freedom are distributed
according to Figure 3, where it can be seen that the nodes of freedom occupy positions that
are different from the geometrical nodes. At every node of freedom there are two variables,
which can be unknown or one of them can be specified as a boundary condition. Inside every
tetrahedron, interior DRM nodes can be added. If interior DRM nodes have not been added,
this code produces 24 equations for every subdomain.
1
2
3
4
5
6
7
8
9
10
Figure 2: Geometrical nodes of a tetrahedral
subdomain.
1 2
3
4
56
Figure 3: Nodes of freedom on a discontinuous
triangular boundary element.
After the assembly is completed, the resulting system of equations is regular. Preliminary
tests with different solvers gave the best performance with an iterative solver using a
Conjugate Gradient-Normal Residual algorithm (Saad, 1996), which is in agreement with the
experience ofPeratta and Popov (2006). This solver with no preconditioning has been used in
the examples presented in this work.
5.2 CODE B: a code using continuous boundary elements
This code uses continuous triangular elements belonging also to the family of elements
proposed by Do Rgo Silva (1994). The nodes of freedom occupy the same position as the
geometrical nodes, hence it produces ten equations for every subdomain, provided interior
DRM nodes have not been added. An important factor must be considered when using
continuous elements. In Section 3 it has been explained that the contribution of a continuous
node towards a closed system of equations cannot be achieved unless the medium is
homogeneous and the plane connections between elements are smooth. When the mesh is
structured, these conditions can be fulfilled, but in a general case the mesh is unstructured,
therefore the condition is not satisfied , unless the node is on the boundary. Consequently,(18) was transformed in such a way that the variables at nodes were potential and partial
-
8/2/2019 13 Mecanica de Fluidos 12
9/20
derivatives instead of potential and normal derivatives. The change of variables has been
done in the following way. Let us apply Eq. (18) to a tetrahedral subdomain and express it in
index notation.
0=
kikj
ijijij qgu
D
ksth with
241
101
101
k
j
i
(19)
Note that the subindex kgoes from 1 to 24 because there are 24 normal derivatives in a
single tetrahedron. At the same time, every normal derivative is the scalar product of the
gradient and the unit normal vector:
kkke
k
e
k
e
k
k ne
un
e
un
e
uq
321
321
+
+
= (20)
By replacing (20) into (19), one obtains
0321
321=
k
eik
k
eik
k
eikjij
ijije
ung
e
ung
e
ungu
D
ksth
kkk(21)
The products between the elements of the G matrix and the components of the unit normal
vectors can be assembled in three 1010 matrices that we will denote 1e
G , 2e
G and 3e
G . The
resulting system of equations is
0
321
321 =
e
ug
e
ug
e
ugu
D
ksth
jeij
jeij
jeijj
ijijij (22)
A further condition arises in order to apply (22): it must be at least four subdomains around
every node in order to have a closed system of equations. This condition is not always
fulfilled in apex nodes that are located in the boundary of the domain, where, instead, at least
one condition of coplanarity can always be found. In those nodes the problem is defined as in
the classical BEM, being the variables the potential and the normal derivative.
The resulting system of equations of the CODE B is over-determined. The over determined
system is solved in a least-squares sense by means of the LSQR algorithm. In the examples
presented here preconditioning has not been used. LSQR is an iterative method for computing
a solutionp to equationsAp = m, whereA is a real matrix with m rows and n columns and m
is a real vector. Developed by Paige and Saunders (1982), it is based on the bidiagonalizationprocedure of Golub and Kahan. It is analytically equivalent to the standard method of
conjugate gradients, but possesses more favourable numerical properties.
6 NUMERICAL EXAMPLES
Preliminary tests on Laplace problems showed maximum errors of 10-8
and 10-11
% for the
codes using discontinuous and continuous elements, respectively. These results are not shown
here since they do not have a domain integral, which would require the use of the DRM for its
solution.
The cases that will be shown correspond to a case of advection diffusion in a prismatic
domain of lengthL = 1 in the1
e direction and width W= 0.2 in the2
e and3
e directions, see
Figure 4. The governing equation is given by (14), whose DRM formulation is given by (18).
The following boundary conditions were applied:
-
8/2/2019 13 Mecanica de Fluidos 12
10/20
u(0, 2e , 3e )=U0 = 10 ; u(L, 2e , 3e )=U1 = 4 (23)
and
2/2 Wenu
= =
2/2 Wenu
= =
2/3 Wenu
= =
2/3 Wenu
= = 0 (24)
L=1.0
W=0.2
W=0.2
e1
e2
e3
Figure 4: Geometry of the domain and the boundary conditions used in the numerical example.
The flow is in the 1e direction, and the flow velocity field is a function of the reaction
constant, k, and the 1e coordinate, and is given as
+
=
2ln
1),( 1
0
11
Lek
U
U
LkeV (25)
For this geometry, velocity field and boundary conditions, the concentration field is
symmetrical to the longitudinal axis, and the results computed on any line parallel to the 1e
axis are the same to the 1D problem with the following known analytical solution:
+
= 1
0
121
012
ln1
2exp),( e
kL
U
U
L
keUkeu (26)
Figure 5 shows the distribution of flow velocities and Figure 6 the analytical solution along
the domain for different values ofk. Two meshes have been used: a rather coarse mesh of 173
subdomains and a finer one of 1456 subdomains (see Figure 7).
6.1 Tests on the scale of the problem
An important issue in this kind of codes was analysed, which is the scale of the problem. A
given problem can be converted into a similar one but with a different scale if a scale factor is
properly applied on both the boundary conditions and the parameters of the governing
equations. Once a result corresponding to this new problem is obtained, using again the same
scale factor the solution of the original problem can be retrieved. From a mathematical point
of view the results should be identical no matter whether scaling has been applied or not. In
the codes presented here the scale affects the time needed for the iterative solver to converge.
The reason is that in the BEM formulation some of the matrices have different dimensions.
For instance, let us consider the resulting system of equations of the advectiondiffusion
-
8/2/2019 13 Mecanica de Fluidos 12
11/20
problem described by (18). The term
k
D
S-T-H has no dimensions, but the matrix G has
a dimension of length, consequently, the size of the problem affects the condition number of
the resulting system of equations. As the DRM formulation is applied locally to everysubdomain, a change in the size of the mesh, that is, when refining the mesh, will change the
condition number.
Velocities
-25
-15
-5
5
15
25
0 0.2 0.4 0.6 0.8 1
k=40
k=20
k=5
k=0
e 1
Figure 5: Magnitude of velocities along the domain for
different values of k.
CONCENTRATIONS
0
2
4
6
8
10
12
0 0.2 0.4 0.6 0.8 1
k=40
k=20
k=5
k=0
e1
Figure 6: Analytical solution of the test cases.
Figure 7: View of the subdivisions of the domain used in the examples.
Figure 8 shows the variation of the condition number with the inverse of the scale factor,
when using both CODES A and B and the mesh of 173 subdomains. The inverse of the scale
factor indicates how many times the characteristic length of the problem has been enlarged. It
can be seen that the condition number has a minimum for an inverse of the scale factor around
two for CODE A, and around sixteen for CODE B. Note that CODE A is better conditioned
than CODE B. Higher values of the inverse of the scale factor cause changes in the condition
number that are more sudden for CODE A and depend on the value of the reaction constant k.
Let us recall that in this problem the velocity field is a function ofk(see Figure 5); the higher
the value ofkthe more advective the problem. As both, the advective term and the reactive
-
8/2/2019 13 Mecanica de Fluidos 12
12/20
term are contained in the term
k
D
S-T-H , this term becomes dominant when k increases;
and as it has no dimensions, changes in the scale have less influence. For instance, for k= 0
the rate of increment of the condition number with the inverse of the scale factor is higherthan fork= 5 and k= 20.
CODE A
(Discontinuous elements)
CODE B
(Continuous elements)
0
500
1000
1500
2000
2500
3000
3500
4000
0 5 10 15 20
k=20
k=5
k=0
Condition number
1/scale factor
0
20000
40000
60000
80000
100000
120000
140000
160000
180000
200000
0 10 20 30
k=20
k=5
k=0
1/scale factor
Condition number
Figure 8: Variation of the condition number with the inverse of the scale factor for both the discontinuous
element code and the continuous element code.
CODE A(Discontinuous elements)
CODE B(Continuous elements)
0
1
2
3
4
5
6
7
8
9
10
0 5 10 15 20
k=20
k=5
k=0
Maximum error [%]
1/scale factor
0
0.5
1
1.5
2
2.5
0 10 20 30
k=20
k=5
k=0
1/scale factor
Maximum error [%]
Figure 9: Variation of the maximum error with the inverse of the scale factor for both the discontinuous element
code and the continuous element code (the values corresponding to the continuous element code with k= 20 are
not displayed because they stay uniformly around 14%).
Figure 9 shows the variation of the maximum error with the inverse of the scale factor. For
CODE A the best accuracy is in the range where the condition number is minimum. For
CODE B the situation is different depending on the value of k. When the term
-
8/2/2019 13 Mecanica de Fluidos 12
13/20
k
D
S-T-H is dominant, the accuracy is not noticeably affected by the scaling, but when it
is small or null, the error reaches a minimum for a scale factor that it is not the best from the
point of view of the condition number. At the same time these optimal scales vary with thevalue ofk.
It is evident that it is convenient to scale the problem when using these codes. A single
change in the unit system in which a given problem is being specified or a refinement of the
mesh can cause the code to fail if scaling is not implemented.
Table 1 and Table 2 compare the time that the solver needs to converge to the solution of
the problem and the maximum error for both situations: with and without scaling, when using
the refined mesh of 1456 subdomains. A better accuracy was expected with regard to the
previous cases, in which a mesh of 173 subdomains was used. Note that in order to obtain a
suitable scale factor it is the size of the subdomain what must be controlled rather than the
size of the whole domain, hence, the scaling criterion was that the average size of thesubdomains had to be kept the same as in the 173 subdomains case when the inverse of the
scale factor was two, for CODE A, and fifteen, for CODE B.
The results displayed in Table 1 and Table 2 show how the rate of convergence of the
solver improves with the scaling. It reduces roughly three times the CPU time for CODE A
and ten times for CODE B. The accuracy worsens when kis equal to five and zero, however,
for these values ofkthe codes achieved good accuracy so the error was low. For the CODE B
and k = 40 the accuracy improves substantially with scaling. The refinement of the mesh
clearly reduces the error when k= 20 but it does not when k is equal to five and zero. The
convergence of the method with the refinement of the mesh shall be analysed in the next
section.
CODE A
k= 40 k= 20 k= 5 k= 0
Not scaled Scaled Not scaled Scaled Not scaled Scaled Not scaled Scaled
Solver time
[s]
2366 746 2666 823 2928 895 3005 923
Maximum
error [%]
6.3 6.5 0.94 0.92 0.039 0.36 0.0078 0.52
Table 1: CODE A (discontinuous elements). Time needed by the iterative solver to reach convergence and the
maximum error with and without scaling, for a refined mesh of 1456 subdomains.
CODE B
k= 40 k= 20 k= 5 k= 0
Not scaled Scaled Not scaled Scaled Not scaled Scaled Not scaled Scaled
Solver time
[s]
1451 103 1503 145 1621 158 1661 169
Maximum
error [%]
33.2 23.0 2.0 1.9 0.13 0.78 0.066 1.0
Table 2: CODE B (continuous elements). Time needed by the iterative solver to obtain a solution and maximum
error with and without scaling, for a refined mesh of 1456 subdomains.
-
8/2/2019 13 Mecanica de Fluidos 12
14/20
6.2 Mesh refinement and interior DRM nodes
Both codes were tested using interior DRM nodes. Two situations were considered: (a)
when one interior DRM node was added in the mesh in the middle of every subdomain, and
(b) when five interior DRM nodes were added, preserving as much as possible equal distancebetween the nodes in order to avoid ill conditioned system of equations. The distribution of
the interior DRM nodes is done automatically by the code, once the number of the interior
DRM nodes is defined in the input data.
It was expected that the accuracy would improve as more interior DRM nodes were added.
Table 3 and Table 4 show that this was the case for both CODES A and B, using a mesh of
173 subdomains. Table 5 shows that this is still the case for the CODE A when using a
refined mesh of 1456 subdomains, but Table 6 shows that not always using interior DRM
nodes a better accuracy for the case of CODE B is achieved. Forkequal to 5 and 0, when
adding one interior DRM node, the accuracy improves in regard with the case with no interior
node; but using 5 interior nodes slightly worsen the accuracy in regard with the case with justone interior node. At the same time, by comparing Table 3 and Table 5 and Table 4 and Table
6 it can be seen that both codes do not converge with mesh refinement when k= 5 and 0.
In order to understand the results of Table 3Table 6, let us recall that the BEM
formulation (7) uses the fundamental solution of the Laplace operator, hence, if the problem
under study is dominated by the Laplace operator, refining the mesh or adding internal DRM
nodes would not necesarily improve the accuracy since new equations are being added, which
worsen the condition number of the resulting system of equations. In the advectiondiffusion
problem considered here, the diffusive process is represented by the Laplacian while the
advective transport is represented by the non-homogeneous term. Under these conditions,
when mesh refinement is applied, in an advection dominated problem the accuracy always
improves because at every subdomain the relative importance of the advective transport isreduced, or to put it another way, the local Peclet (Pe) number is reduced, since if a length
scale linked to the size of the subdomain is used to calculate the Pe number at every
subdomain, every time the mesh is refined, the Pe number decreases. If after refining the
mesh the advection is still relevant, the accuracy can be improved by adding internal DRM
nodes.
CODE A
k= 40 k= 20 k= 5 k= 0
Number of
internal
DRM nodes0 1 5 0 1 5 0 1 5 0 1 5
Solver time
[s]
97 107 141 95 102 136 102 108 144 104 111 145
Maximum
error [%]
29.6 28.3 23.3 4.37 4.14 3.12 0.17 0.16 0.13 0.03 0.03 0.03
Table 3: CODE A (discontinuous elements). Maximum error with various number of interior DRM nodes. Mesh
of 173 subdomains.
-
8/2/2019 13 Mecanica de Fluidos 12
15/20
CODE B
k= 40 k= 20 k= 5 k= 0
Number of
internalDRM nodes
0 1 5 0 1 5 0 1 5 0 1 5
Solver time
[s]
10 17 45 10 19 47 12 20 55 13 24 63
Maximum
error [%]
120 99 77 14.4 12.0 9.2 0.63 0.43 0.33 0.99 0.65 0.34
Table 4: CODE B (continuous elements). Time needed by the solver to obtain a solution and maximum error
with various number of interior DRM nodes. Mesh of 173 subdomains.
CODE A
k= 40 k= 20 k= 5 k= 0
Number ofinternal
DRM nodes
0 1 5 0 1 5 0 1 5 0 1 5
Solver time
[s]
746 802 1115 823 885 1220 895 959 1304 923 986 1350
Maximum
error [%]
6.5 6.08 4.41 0.92 0.90 0.67 0.36 0.24 0.05 0.52 0.36 0.07
Table 5: CODE A (discontinuous elements). Time needed by the solver to obtain a solution and maximum error
with various number of interior DRM nodes, for a refined mesh of 1456 subdomains.
CODE B
k= 40 k= 20 k= 5 k= 0
Number of
internal
DRM nodes0 1 5 0 1 5 0 1 5 0 1 5
Solver time
[s]
103 201 637 145 231 707 158 306 972 169 325 1027
Maximum
error [%]
23.0 26.3 30.1 1.9 1.61 1.37 0.78 0.51 0.67 1.0 0.65 0.71
Table 6: CODE B (continuous elements). Time needed by the solver to obtain a solution and maximum error
with various number of interior DRM nodes, for a refined mesh of 1456 subdomains.
The apex of the subdomains of the coarser mesh is about 1/10 long, and the one of the finermesh is about 1/17 long. D was set equal to 1 in all the examples. Figure 10 and Figure 11
show the values of the local Pe along the subdomain. Figure 10 shows that the local Pe is
below 0.34 and 0.09 forkequal to 5 and 0, respectively, when using the coarser mesh. As the
transport is mainly diffusive along the whole domain, the accuracy would not improve by
refining the mesh. And this is what has been observed in both CODES A and B. Under these
conditions, the results in Table 4 and Table 6 show that internal DRM nodes are not of much
help for CODE B. Table 3 and Table 5 show that adding internal DRM nodes can improve the
accuracy of CODE A.
-
8/2/2019 13 Mecanica de Fluidos 12
16/20
Pe number
Mesh 173
0.0
0.5
1.0
1.5
2.0
2.5
0 0.2 0.4 0.6 0.8 1
k=40
k=20
k=5
k=0
e1
Figure 10: Local Peclet number distribution for the
coarser mesh.
Pe number
Mesh 1456
0.0
0.5
1.0
1.5
2.0
2.5
0 0.2 0.4 0.6 0.8 1
k=40
k=20
k=5
k=0
e1
Figure 11: Local Peclet number distribution for the
finer mesh.
When kis equal to 20, in vast parts of the domain the problem is advective enough as to
show convergence with mesh refinement. When kis above 20, adding DRM nodes improves
the accuracy for the same reason, and although at a first sight the results of the CODE B seem
not to be in agreement with this conclusion, it shall be seen that they are not an exception.
Figure 12 and Figure 13 show the distribution of the absolute and relative error when k=
40 and the problem is solved with CODE B. The crosses correspond to NO internal DRM
nodes, the triangles correspond to 1 internal DRM node and the dots to 5 internal DRM
nodes. Figure 14 and Figure 15 show the same for CODE A. Figure 12 shows that adding
internal DRM nodes in the case of CODE B reduces the absolute error where it is higher: in
the zone where advection is more important, and slightly increases it in the middle of the
domain, where diffusion is dominant. In the advective part of the domain, the absolute error istwo orders of magnitude higher than in the middle. But, it is in the middle of the domain
where the relative error is higher because the concentrations are the lowest, and that is why
the maximum relative error increases with internal DRM nodes in Figure 13. Conversely,
Figure 15 shows that in the CODE A case the relative error is distributed more evenly, then,
adding internal DRM nodes improves both absolute and relative errors.
-0.35
-0.25
-0.15
-0.05
0.05
0.0 0.2 0.4 0.6 0.8 1.0
No int. node
1 int. node
5 int. node
e 1
Absolute error
Figure 12: CODE B. Absolute error distribution fork
= 40.
-20
-10
0
10
20
30
40
0.0 0.2 0.4 0.6 0.8 1.0
No int. node
1 int. node
5 int. node
e 1
Relative error [%]
Figure 13: CODE B: Relative error distribution fork=
40.
According to the previous analysis, internal DRM nodes can be used to improve the
accuracy but they should be placed only in those subdomains where advection is important.
For instance, the test problem was solved with CODES A and B , with k= 40 and using thefiner mesh with interior DRM nodes only in those subdomains where Pe > 0.65. Table 7
-
8/2/2019 13 Mecanica de Fluidos 12
17/20
compares the obtained results with the ones of the codes using DRM nodes in every
subdomain. Results of both codes show that similar accuracy is achieved in both situation,
that is, when using internal nodes inside every subdomain and when using internal nodes only
where Pe > 0.65; but in the later situation a good deal of solver time is gained compared withthe former one.
-0.16
-0.12
-0.08
-0.04
0.00
0.04
0.0 0.2 0.4 0.6 0.8 1.0
No int. node
1 int. node
5 int. node
e 1
Absolute error
Figure 14: CODE A. Absolute error distribution fork=
40.
-8
-6
-4
-2
0
2
0.0 0.2 0.4 0.6 0.8 1.0
No int. node
1 Int. node
5 Int. node
e 1
Relative error [%]
Figure 15: CODE A: Relative error distribution fork=
40.
CODE A - k= 40 CODE B - k= 40
Internal DRM
nodes in every
subdomain
Internal DRM
nodes only where
Pe > 0.65
Internal DRM nodes in
every subdomain
Internal DRM nodes only
where Pe > 0.65
1
IDRMN
5
IDRMN
1
IDRMN
5
IDRMN1 IDRMN 5 IDRMN 1 IDRMN 5 IDRMN
Solver time[s] 802 1115 777 927 201 637 174 381
Maximum
error [%]
6.08 4.41 6.11 4.53 26.3 30.1 25.9 29.0
Table 7: Comparison of the performance of the both codes using internal DRM nodes in every subdomain and
only where Pe > 0.65.
6.3 Comparison of performance of CODES A and B
Table 3Table 6 show that as long as the same mesh is used, CODE A is more accurate
than CODE B, specially when the problem is dominated by advection, but the time required
by CODE A to reach the solution is one order of magnitude higher than the time needed byCODE B. Let us recall that CODE A produces 24 equations for every subdomain while
CODE B produces 10, provided no internal DRM nodes are added. In order to compare the
performance of both codes having a similar number of degrees of freedom, a mesh of 413
subdomains was produced and the example was solved with code B, with no internal DRM
nodes (Table 8). In this case, CODE B produces a system of equation of 4130 degrees of
freedom, which is very close to the 4152 degrees of freedom of the system produced by
CODE A when using the 173 subdomain mesh (Table 3). If the comparison is done in this
way, the difference between both codes regarding both accuracy and solver time is reduced,
being the reduction more important as the problem is more advective. However, the trend is
the same as it was observed previously, the CODE A shows to be more accurate and slower
than CODE B.Note that the time appearing in Table 1Table 8 corresponds only to the time needed by
-
8/2/2019 13 Mecanica de Fluidos 12
18/20
the solver to converge to a solution; it does not compute the time required by other tasks of
the codes such as the assembling of the system of equation. Let us recall that CODE A
produces a regular system of equations while CODE B produces an overdetermined system,
hence they use different solvers (Section 5). There is no simple argument to explain themagnitude of the difference observed between the solvers time. For example, the efficiency of
the solvers is influenced by factors such as the condition number and the structure of the
matrix of coefficients of the system. The two codes produce different matrices, which affects
the efficiency of the solvers. A proper answer requires further research, similar to the one
already completed for 2D DRM-MD codes (see Portapila and Power, 2001, 2005), which is
currently being carried out.
CODE B
k= 40 k= 20 k= 5 k= 0
Solver time [s] 28 29 35 37
Maximum error [%] 46.2 5.52 0.73 1.01
Table 8: Solver time and maximum error for CODE B using a mesh of 413 subdomains.
7 CONCLUSIONS
In this paper, the sensitivity of the 3D implementation of the boundary element dual
reciprocity methodmultidomain approach (DRM-MD) to the continuity of the elements,
scaling, mesh refinement and number of internal nodes, has been tested in an advection
diffusion problem with non-uniform velocity field.
It has been shown that implementation of scaling is essential for these kind of problems.
The examples presented here are sensitive to length scaling, but it must be noted that when
changing the governing equation other scales can become relevant. Tests were made on an
advectiondiffussion problem with non-uniform velocity field using both discontinuous and
continuous elements (CODES A and B, respectively). The size of the subdomains, which
depends on the size of the problem and the degree of refinement of the mesh, affects the
condition number of the resulting system of equations and consequently the time needed by
the iterative solvers to converge. By using scaling of the equations, the solver time was
reduced by three and by ten times for the CODES A and B, respectively.
For the examples tested, interior DRM nodes in every subdomain improved the accuracy
of both codes provided they are used in subdomains where the advective transport, that is, thenon-homogeneous term, plays a noticeable role, for instance, where the local Pe is above 0.5.
Adding interior DRM nodes where the local Pe is lower than 0.5 would only increase the
CPU time required by the iterative solvers without improving or even worsening the
accuracy. The same can be said about mesh refinement. Every time the mesh is refined the
DRM formulation becomes more diffusive at the level of every subdomain, and as the
diffusive term is computed by the Laplacian, it does not make sense to refine the mesh further
in areas where the local Pe is below 0.5.
Regarding the use of continuous and discontinuous elements, it is difficult to say which
strategy is better. The discontinuous element code offers higher accuracy, especially for
highly advective transport, but it is much slower than the continuous element code. At the
present stage, the discontinuous element code has the advantage that it can solve non-homogeneous domains. The most convenient strategy is to have a choice to use both types of
-
8/2/2019 13 Mecanica de Fluidos 12
19/20
elements in a single mesh, which would preserve the versatility of dealing with non-
homogeneous domains, while offering possibility for CPU and memory usage reduction.
REFERENCESBuhmann, M.D., A new class of radial basis function with compact support. Math Comput,
70/233:30718, 2001.
Do Rgo Silva, J.J., Acoustic and elastic wave scattering using boundary elements.
Southampton and Boston: Computational Mechanics Publications, 1994.
Floater, M.S., and Iske, A., Multistep scattered data interpolation using compactly supported
radial basis functions.J Comput Appl Math, 73:6578, 1996.
Florez, W.F.,Nonlinear flow using dual reciprocity. Southampton: WIT Press, 2001.
Florez, W.F., and Power, H., DRM multidomain mass conservative interpolation approach for
the BEM solution of the two-dimensional NavierStokes equations. Comput Math Appl,
43/3-5:45772, 2002a.
Florez, W.F, and Power, H., Multi-domain mass conservative dual reciprocity method for the
solution of the non-Newtonian Stokes equations.Appl Math Model, 26/3:397419, 2002b.
Golberg, M.A., and Chen, C.S., The theory of radial basis functions applied to the BEM for
inhomogeneous partial differential equations.Bound Elem Commun, 5:5761, 1994.
Golberg, M.A., and Chen, C.S., A bibliography on radial basis functions approximation.
Bound Elem Commun, 7:15563, 1996.
Nardini, D., and Brebbia, C.A., A new approach to free vibration analysis using boundary
elements.Appl Math Model, 7:15762, 1983.
Natalini, B., The boundary element dual reciprocity methodmultidomain approach for
solving 3D potential problems. Ph.D. thesis, University of Wales, 2005.
Natalini, B., and Popov, V., Radial basis function selection in 3D DRMMD for potentialproblems. In: Brebbia CA, editor. Boundary elements XXVI. Southampton: WIT Press,
10110, 2004.
Natalini, B., and Popov, V., Tests of radial basis functions in the 3D DRMMD. Commun
Numer Methods Eng, 22:1322, 2006.
Paige, C.C., and Saunders, M.A., Algorithm 583: LSQR: sparse linear equations and least
squares problems.ACM Trans Math Software, 8/2:195209, 1982.
Peratta, A., and Popov, V., A new scheme for numerical modelling of flow and transport
processes in 3D fractured porous media.Adv Water Resour, 29:4261, 2006.
Popov, V., and Power, H., DRM-MD approach for the numerical solution of gas flow in
porous media with application to landfill.Eng Anal Bound Elem, 23:17588, 1999a.
Popov, V., and Power H., The DRM-MD integral equation method: an efficient approach forthe numerical solution of domain dominant problems. Int J Numer Meth Eng, 44:32753,
1999b.
Popov, V., and Power, H., Numerical analysis of the efficiency of landfill venting trenches. J
Environ Eng ASCE, 126/1:328, 2000.
Popov, V., Power, H., and Baldasano, J.M., BEM solution of design of trenches in a multi-
layered landfill.J Environ Eng ASCE, 124/1:5966, 1998.
Portapila, M.I., and Power, H., Efficiency and accuracy of an iterative solution for the DRM-
MD.Bound Elem Commun, 12/1:2637, 2001.
Portapila, M.I., and Power H., Iterative schemes for the solution of system of equations
arising from the DRM in multi-domain approach, and a comparative analysis of theperformance of two different radial basis functions used in the interpolation. Eng Anal
Bound Elem, 29/2:10725, 2005.
-
8/2/2019 13 Mecanica de Fluidos 12
20/20
Rodrguez, J.J., and Power, H., An adaptative dual reciprocity scheme for the numerical
solution of the Poisson equation.Eng Anal Bound Elem, 26:283300, 2002.
Saad, Y.,Iterative methods for sparse linear systems. New York: PWS Publishing, 1996.
Schaback, R., Creating surfaces from scattered data using radial basis functions. In: Lyche,T., Dhlen, M., Schumaker, L., editors. Mathematical methods for curves and surfaces.
Nashville: Vanderbilt University Press, 47796, 1995.
Schclar, N.A., Application of the boundary element method to the structural analysis of three
dimensional anisotropic material. Ph.D. thesis, University of Portsmouth, 1993.
Wendland, H., Piecewise polynomial, positive definite and compactly supported radial
function of minimal degree.Adv Comput Math, 4:38996, 1995.
Wu, Z., Multivariate compactly supported positive definite radial functions. Adv Comput
Math, 4:28392, 1995.
top related